ON THE GEOMETRY OF PROJECTIVE TENSOR PRODUCTS
OHAD GILADI, JOSCHA PROCHNO, CARSTEN SCHUTT,¨ NICOLE TOMCZAK-JAEGERMANN, AND ELISABETH WERNER
n Abstract. In this work, we study the volume ratio of the projective tensor products `p ⊗π n n `q ⊗π `r with 1 ≤ p ≤ q ≤ r ≤ ∞. We obtain asymptotic formulas that are sharp in almost all cases. As a consequence of our estimates, these spaces allow for a nearly Euclidean decomposition of Kaˇsintype whenever 1 ≤ p ≤ q ≤ r ≤ 2 or 1 ≤ p ≤ 2 ≤ r ≤ ∞ and q = 2. Also, from the Bourgain-Milman bound on the volume ratio of Banach spaces in terms of their cotype 2 constant, we obtain information on the cotype of these 3-fold projective tensor products. Our results naturally generalize to k-fold products `n ⊗ · · · ⊗ `n with k ∈ p1 π π pk N and 1 ≤ p1 ≤ · · · ≤ pk ≤ ∞.
1. Introduction In the geometry of Banach spaces the volume ratio vr(X) of an n-dimensional normed space X is defined as the n-th root of the volume of the unit ball in X divided by the volume of its John ellipsoid. This notion plays an important role in the local theory of Banach spaces and has significant applications in approximation theory. It formally originates in the works [Sza78] and [STJ80], which were influenced by the famous paper of B. Kaˇsin[Kaˇs77] on nearly Euclidean orthogonal decompositions. Kaˇsindiscovered that for arbitrary n ∈ N, 2n the space `1 contains two orthogonal subspaces which are nearly Euclidean, meaning that n their Banach-Mazur distance to `2 is bounded by an absolute constant. S. Szarek [Sza78] n noticed that the proof of this result depends solely on the fact that `1 has a bounded volume n ratio with respect to `2 . In fact, it is essentially contained in the work of Szarek that if X is a 2n-dimensional Banach space, then there exist two n-dimensional subspaces each n having a Banach-Mazur distance to `2 bounded by a constant times the volume ratio of X squared. This observation by S. Szarek and N. Tomczak-Jaegermann was further investigated in [STJ80], where the concept of volume ratio was formally introduced, its connection to the cotype 2 constant of Banach spaces was studied, and Kaˇsintype decompositions were proved n n for some classes of Banach spaces, such as the projective tensor product spaces `p ⊗π `2 , 1 ≤ p ≤ 2. Given two vector spaces X and Y , their algebraic tensor product X ⊗ Y is the subspace of the dual space of all bilinear maps on X × Y spanned by elementary tensors x ⊗ y, x ∈ X, y ∈ Y (a formal definition is provided below). The theory of tensor products was established by A. Grothendieck in 1953 in his R´esum´e[Gro53] and has a huge impact on
Date: April 6, 2017. 2010 Mathematics Subject Classification. 46A32, 46B28. O.G. was supported in part by the Australian Research Council. J.P. was supported in part by the FWF grant FWFM 162800. A part of this work was done when N.T-J. held the Canada Research Chair in Geometric Analysis; also supported by NSERC Discovery Grant. E.W. was supported by NSF grant DMS-1504701.
1 Banach space theory (see, e.g., the survey paper [Pis12]). This impact and the success of the concept of tensor products is to a large extent due to the work [LP68] of J. Lindenstrauss and A. Pe lczynskiin the late sixties who reformulated Grothendieck’s ideas in the context of operator ideals and made this theory accessible to a broader audience. Today, tensor products appear naturally in numerous applications, among others, in the entanglement of qubits in quantum computing, in quantum information theory in terms of (random) quantum channels (e.g., [AS06,ASW10,ASW11,SWZ11]) or in theoretical computer science to represent locally decodable codes [Efr09]. For an interesting and recently discovered connection between the latter and the geometry of Banach spaces we refer the reader to [BNR12]. The geometry of tensor products of Banach spaces is complicated, even if the spaces involved are of simple geometric structure. For example, the 2-fold projective tensor product of Hilbert spaces, `2 ⊗π `2, is naturally identified with the Schatten trace class√S1, the space ∗ of all compact operators T : `2 → `2 equipped with the norm kT kS1 = trace T T . This space does not have local unconditional structure [GL74]. The geometric structure of triple tensor products is even more complicated and therefore it is hardly surprising that very little is known about the geometric properties of these spaces. For instance, regarding the permanence of cotype (see below for the definition) under projective tensor products, it was proved by N. Tomczak-Jaegermann in [TJ74] that the space `2 ⊗π `2 has cotype 2, but the corresponding question in the 3-fold case is still open for more than 40 years. G. Pisier proved in [Pis90] and [Pis92] that the space Lp ⊗π Lq has cotype max (p, q) if p, q ∈ [2, ∞) and, till the present day, it is unknown whether these spaces have a non-trivial cotype when p ∈ (1, 2) and q ∈ (1, 2]. In the recent paper [BNR12] by J. Bri¨et,A. Naor and O. Regev 1 1 1 they showed that the spaces `p ⊗π `q ⊗π `r fail to have non-trivial cotype if p + q + r ≤ 1. Their proof uses deep results from the theory of locally decodable codes. A direct, rather surprising consequence of their work is that for p ∈ (1, ∞) the space `p ⊗π `2p/(p−1) ⊗π `2p/(p−1) fails to have non-trivial cotype, while, by Pisier’s result, the 2-fold projective tensor product `2p/(p−1) ⊗π `2p/(p−1) has finite cotype. Let us also mention that interest in cotype is, to a great deal, due to a famous result of B. Maurey and G. Pisier [MP76], who showed that a n Banach space X fails to have finite cotype if and only if it contains `∞’s uniformly. In view of the various open questions and surprising results around the geometry of projec- tive tensor products, with the present paper, we contribute to a better understanding of the n n n geometric structure of 3-fold projective tensor products `p ⊗π `q ⊗π `r (1 ≤ p ≤ q ≤ r ≤ ∞) by studying one of the key notions in local Banach space geometry, the volume ratio of these spaces. This provides new structural insight and allows, on the one hand, to draw conclusions regarding cotype properties of these spaces and, on the other hand, to see which of these spaces allow a nearly Euclidean decomposition of Kaˇsintype.
2. Presentation of the main result
Given two Banach spaces, (X, k · kX ) and (Y, k · kY ), the projective tensor product space, denoted X ⊗π Y , is the space X ⊗ Y equipped with the norm ( m m ) X X kAkX⊗πY = inf kxikX kyikY : A = xi ⊗ yi, xi ∈ X, yi ∈ Y . (2.1) i=1 i=1 See Section 3.3 below for more information on projective tensor products.
2 Given an n-dimensional Banach space (X, k · kX ), let BX denote its unit ball, and let EX be the ellipsoid of maximal volume contained in BX . The volume ratio of X is defined by vol (B )1/n vr(X) = n X , (2.2) voln(EX )
where voln(·) denotes the n-dimensional Lebesgue measure. Considering the importance of tensor products, it is natural to study the geometric prop- erties of X ⊗π Y and, in particular, the volume ratio vr(X ⊗π Y ) of these spaces. As already n n mentioned in the introduction, when X = `p (1 ≤ p ≤ 2) and Y = `2 this was carried out in [STJ80]. The complete answer was given later by C. Sch¨uttin [Sch82], where it was proved that if 1 ≤ p ≤ q ≤ ∞, then 1, q ≤ 2, 1 − 1 n 2 q , p ≤ 2 ≤ q, 1 + 1 ≥ 1, n n p q vr `p ⊗π `q p,q 1 − 1 n p 2 , p ≤ 2 ≤ q, 1 + 1 ≤ 1, p q max 1 − 1 − 1 ,0 n ( 2 p q ), p ≥ 2.
The notation p,q means equivalence up to constants that depend only on p and q. In [DM05] this was generalized by A. Defant and C. Michels to the setting E ⊗π F , where E and F are symmetric Banach sequence spaces, each either 2-convex or 2-concave. n n n We study tensor products `p ⊗π `q ⊗π `r (1 ≤ p ≤ q ≤ r ≤ ∞). Our main result is as follows.
Theorem A. Let n ∈ N and 1 ≤ p ≤ q ≤ r ≤ ∞. Then we have 1, r ≤ 2, 1 1 1 max( 2 − q − r ,0) 1 1 1 n n n n , p ≤ 2 ≤ q, p + q + r ≥ 1, vr `p ⊗π `q ⊗π `r p,q,r 1 − 1 n p 2 , p ≤ 2 ≤ q, 1 + 1 + 1 ≤ 1, p q r max 1 − 1 − 1 − 1 ,0 n ( 2 p q r ), p ≥ 2. In the case p ≤ q ≤ 2 ≤ r, we have
1 1 1 1 n n n min( 2 − r , q − 2 ) 1 .p,q,r vr `p ⊗π `q ⊗π `r .p,q,r n .
Here and in what follows .p,q,r, &p,q,r mean inequalities with implied positive constants that depend only on the parameters p, q, r. The asymptotic notation p,q,r means that we have both .p,q,r and &p,q,r. Remark 2.1. We would like to remark that whenever 1 ≤ p ≤ q ≤ 2 ≤ r ≤ ∞, we are able to improve on the general bound given by TheoremA in the following situations (see Corollary 5.7):
min 1 − 1 , 1 − 1 n ( q 2 2 r ), p = 1 ≤ q ≤ 2 ≤ r, n n n 1 − 1 vr `p ⊗π `q ⊗π `r p,q,r n q 2 , p = q ≤ 2, r = ∞, 1, 1 ≤ p ≤ q = 2 ≤ r.
3 n n n Remark 2.2. TheoremA and Remark 2.1 immediately imply that the spaces `p ⊗π `q ⊗π `r allow a nearly Euclidean decomposition of Kaˇsintype, when 1 ≤ p ≤ q ≤ r ≤ 2 or 1 ≤ p ≤ 2 ≤ r ≤ ∞ and q = 2. Before we proceed, let us comment on the strategy of the proof. Recall first that a Banach space (X, k · kX ) is said to have enough symmetries if the only operators that commute with every isometry on X are multiples of the identity. It is known that if (X, k · kX ) is n-dimensional and has enough symmetries, then EX is given by n −1 n EX = kid : `2 → Xk B2 , n where B2 denotes the n-dimensional Euclidean ball (see, for instance, [TJ89, Sec. 16]). Hence, using formula (2.2), if (X, k · kX ) is n-dimensional and has enough symmetries, then 1/n voln(BX ) n (∗) √ 1/n n vr(X) = n id : `2 → X n voln(BX ) id : `2 → X , (2.3) voln(B2 ) n 1/n √ where in (∗) we used Stirling’s formula to deduce vol(B2 ) 1/ n. It is also known that projective tensor products of `p spaces are spaces with enough symmetries [Sch82, STJ80] n n n and therefore formula (2.3) holds for the spaces `p ⊗π `q ⊗π `r . Thus, in order to prove n n n TheoremA, it is enough to compute the volume of the unit ball of `p ⊗π `q ⊗π `r , and the n3 n n n norm of the natural identity between `2 and `p ⊗π `q ⊗π `r . The rest of this paper is organized as follows. In Section3 we collect some basic results n n n which will be used later. In Section4 we estimate the volume of the unit ball in `p ⊗π `q ⊗π `r . In Section5 we estimate the norm of the natural identity. In Section6 we discuss the case of k-fold tensor products. Finally, in Section7, we present some applications of TheoremA.
3. Preliminaries In this section, we introduce the necessary notions and background material and provide the main ingredients needed to prove the estimates for the volume ratio of 3-fold projective tensor products. These include an extension of Chevet’s inequality, a lower bound on the volume of unit balls in Banach spaces due to Sch¨utt,the famous Blaschke-Santal´oinequality, and a multilinear version of an inequality of Hardy and Littlewood.
3.1. General notation. Given a Banach space (X, k · kX ), denote its unit ball by BX and its dual space by X∗. For two Banach spaces X and Y , we write L(X,Y ) for the space of all bounded linear operators from X to Y . n n For 1 ≤ p ≤ ∞, `p is the vector space R with the norm ( Pn |x |p1/p , 1 ≤ p < ∞ n i=1 i k(xi)i=1kp = max |xi| , p = ∞. 1≤i≤n n n n ∗ The unit ball in `p is denoted by Bp = {x ∈ R : kxkp ≤ 1}. The conjugate p of p is 1 1 defined via the relation p + p∗ = 1. Unless otherwise stated, e1, . . . , en will be the standard unit vectors in Rn. We shall also use the asymptotic notations . and & to indicate the corresponding in- equalities up to universal constant factors, and we shall denote equivalence up to universal
4 constant factors by , where A B is the same as (A . B) ∧ (A & B). If the constants involved depend on a parameter α, we denote this by .α, &α and α, respectively.
3.2. Polar body and Blaschke-Santal´oinequality. A convex body K in Rn is a compact convex subset of Rn with non-empty interior. The n-dimensional Lebesgue measure of a n convex body K ⊆ R is denoted by voln(K). If 0 is an interior point of a convex body K in Rn, we define the polar body of K by ◦ n K := y ∈ R : ∀x ∈ K : hx, yi ≤ 1 , where h·, ·i stands for the standard inner product on Rn. Note that the unit ball of any norm on Rn is a convex body and its polar body is just the unit ball of the corresponding dual norm. Moreover, we have (K◦)◦ = K. The famous Blaschke-Santal´oinequality provides a sharp upper bound for the volume product of a convex body with its polar (see, for example, [Pis89, Sec. 7] or [Sch14]). Lemma 3.1 (Blaschke-Santal´oinequality). Let K be an origin symmetric convex body in Rn. Then ◦ n 2 voln(K) · voln(K ) ≤ voln(B2 ) , with equality if and only if K is an ellipsoid. 3.3. Tensor products and extended Chevet inequality. Given two Banach spaces (X, k · kX ) and (Y, k · kY ), the algebraic tensor product X ⊗ Y can be constructed as the space of linear functionals on the space of all bilinear forms on X × Y . Given x ∈ X, y ∈ Y and a bilinear form B on X × Y , we define (x ⊗ y)(B) := B(x, y). On the tensor product space define the projective tensor product space, denoted by X ⊗π Y , as X ⊗ Y equipped with the norm ( m m ) X X kAkX⊗πY := inf kxikX kyikY : A = xi ⊗ yi, xi ∈ X, yi ∈ Y . i=1 i=1
Also, define the injective tensor norm space, denoted X ⊗ Y , as the tensor product space X ⊗ Y equipped with the norm
( n m ) X X kAk := sup ϕ(x )ψ(y ) : A = x ⊗ y , ϕ ∈ B ∗ , ψ ∈ B ∗ . X⊗Y i i i i X Y i=1 i=1 Here, we only consider finite-dimensional spaces and therefore the tensor products are always ∗ complete. In this case, it can be shown that X ⊗π Y = N (X ,Y ), the space of all nuclear ∗ ∗ operators from X into Y . We will often use the fact that (X ⊗π Y ) , the dual space of ∗ ∗ X ⊗π Y , is the space of operators from X to Y , L(X ,Y ), equipped with the standard operator norm. It is also known that L(X∗,Y ) can be identified with the injective tensor product X ⊗ Y . In particular, for A ∈ X ⊗ Y , we have
kAkX⊗Y = sup kAxkY . x∈BX∗ We refer the reader to [Rya02] and [DFS08] for more information about tensor products.
5 n Recall that for a sequence x = (xi)i=1 in a Banach space (X, k · kX ), the norm kxkω,2 is given by 1 n ! 2 X 2 kxkw,2 := sup |ϕ(xi)| . kϕkX∗ =1 i=1 The following inequality is due to Chevet [Che78].
Lemma 3.2 (Chevet’s inequality). Let (X, k · kX ), (Y, k · kY ) be two Banach spaces and m,n m n consider sequences x1, . . . , xm ∈ X, y1, . . . , yn ∈ Y , and sequences (gi,j)i,j=1, (ξi)i=1, (ηj)j=1 of independent identically distributed standard Gaussians random variables. Then m,n n m X m X n X E gi,jxi ⊗ yj ≤ k(xi)i=1k E ηjyj + (yj)j=1 E ξixi . (3.1) ω,2 ω,2 i,j=1 X⊗Y j=1 Y i=1 X
∗ ∗ ∗ ∗ It is known that BX ⊗πY = conv (BX ⊗ BY ) (see, for example, Proposition 2.2 in [Rya02]) and thus
(xi ⊗ yj)i,j ω,2 = (xi)i ω,2 · (yj)j ω,2. (3.2) Using inequalities (3.1) and (3.2), we obtain the following 3-fold version of Chevet’s inequal- ity.
Lemma 3.3 (3-fold Chevet inequality). Let (X, k · kX ), (Y, k · kY ), (Z, k · kZ ) be Banach spaces. Assume that x1, . . . , xm ∈ X, y1, . . . , yn ∈ Y and z1, . . . , z` ∈ Z. Let gi,j,k, ξi, ηj, ρk, i = 1, . . . , m, j = 1, . . . , n, k = 1, . . . , `, be independent standard Gaussians random variables. Then m,n,` X E gijk xi ⊗ yj ⊗ zk ≤ Λ,
i,j,k=1 X⊗Y ⊗Z where ` n m n X m ` X Λ := k(xi)i=1k (yj)j=1 E ρkzk + k(xi)i=1k (zk)k=1 E ηjyj w,2 w,2 w,2 w,2 k=1 Z j=1 Y m n ` X + (yj)j=1 (zk)k=1 E ξixi . w,2 w,2 i=1 X We would like to point out that, simply using the triangle inequality, a corresponding lower bound can be obtained up to an absolute constant. 3.4. Volume ratio and Rademacher cotype. The concept of Rademacher cotype was introduced to Banach space theory by J. Hoffmann-Jørgensen [HJ74] in the early 1970s. The basic theory was developed by B. Maurey and G. Pisier [MP76]. A Banach space (X, k · kX ), is said to have Rademacher cotype α if there exists a constant C ∈ (0, ∞) such that for all m ∈ N and all x1, . . . , xm ∈ X, m m 1/α X α X kxikX ≤ C E εixi . (3.3) X i=1 i=1 ∞ In (3.3) and in what follows, (εi)i=1 denotes a sequence of independent symmetric Bernoulli 1 random variables, that is, P(εi = 1) = P(εi = −1) = 2 for all i ∈ N. The smallest C
6 that satisfies (3.3) is denoted by Cα(X) and called the cotype α constant of X. By taking x1 = x2 = ··· = xm it follows that necessarily α ≥ 2. 1 1 1 In [BNR12] it was shown that, whenever p + q + r ≤ 1, then
Cα (`p ⊗π `q ⊗π `r) = ∞, for every 2 ≤ α < ∞. However, it is still an open question whether for example `2 ⊗π `2 ⊗π `2 has finite cotype. One reason to study volume ratio of tensor products is its relation to the cotype con- stant. More precisely, given an estimate on volume ratio, one can use the following result by Bourgain and Milman [BM87]) which connects volume ratio and cotype of a Banach space. Theorem 1 ([BM87]). Let X be a Banach space. Then
vr(X) . C2(X) log (2C2(X)) . The relation between volume ratio and cotype property is far from being well understood. For example, it was asked in [STJ80] whether bounded volume ratio implies cotype q for every q > 2. n n n In Section7 below we discuss the cotype property of the space `p ⊗π `q ⊗π `r , as well as the extended case of k-fold tensor products for k > 3. 3.5. Volume of unit balls in Banach spaces. The following result is a special case of Lemma 1.5 in [Sch82]. It provides a lower estimate for the volume of the unit ball of a n normed space by the volume of a B∞ ball of a certain radius, arising from an average over sign vectors.
Lemma 3.4. Let (X, k · kX ) be an n-dimensional Banach space, and let e1, . . . , en be basis ∞ vectors such that keikX = 1, and (εi)i=1 a sequence of independent symmetric Bernoulli random variables. Then n !−n n X 2 E εiei ≤ voln(BX ).
i=1 X We will use this result in combination with the 3-fold version of Chevet’s inequality to n n n obtain a lower bound on the volume of the unit ball in `p∗ ⊗ `q∗ ⊗ `r∗ which gives, using the Blaschke-Santal´oinequality, an upper bound on the volume of the unit ball in the dual space `p ⊗π `q ⊗π `r, as well as tensor products of more than three `p spaces. 3.6. Rademacher versus Gaussian averages. In order to use Chevet’s inequality in combination with Lemma 3.4, we need to pass from a Rademacher average to a Gaussian one. The following result due to Pisier shows that Rademacher averages are dominated by Gaussian averages in arbitrary Banach spaces. Note however that in general these averages are not equivalent.
Lemma 3.5 ([Pis86]). Let (X, k · kX ) be a Banach space, 1 ≤ p < ∞ and let ξ1, . . . , ξn be independent, symmetric random variables. Assume that E|ξi| = E|ξj| for all 1 ≤ i, j ≤ n. Then, for all x1, . . . , xn ∈ X, we have n p n p X −p X E εixi ≤ E|ξ1| E ξixi .
i=1 i=1
7 In particular, if we choose g1, . . . , gn to be independent Gaussian random variables, then −p p/2 since E|g1| = (π/2) , we have n p n p X X E εixi p E gixi . . i=1 i=1 3.7. A multilinear Hardy-Littlewood type inequality. An essential tool in proving n3 n n n upper bounds on the norm of the natural identity between `2 and `p ⊗π `q ⊗π `r is the fol- lowing inequality, which is a generalization of a classical inequality by Hardy and Littlewood [HL34]. For now, we state it only for the case of 3-fold tensors. In Section6 we also present the consequences of the general version.
1 1 1 1 Theorem 2 ([PP81], Thm. B). Let n ∈ N and 1 ≤ p, q, r ≤ ∞ so that p + q + r ≤ 2 . Then, n n n for all A ∈ `p∗ ⊗ `q∗ ⊗ `r∗ , we have
kAk n3 kAk`n ⊗ `n ⊗ `n , `µ . p∗ q∗ r∗ where µ is given by 3 µ := 1 1 1 . (3.4) 2 − p − q − r
n n n 4. The volume of the unit ball in `p ⊗π `q ⊗π `r n n n In this section we evaluate the volume of the unit ball of `p ⊗π `q ⊗π `r up to constants depending only on the parameters p, q, r. The main ingredients in the proof are the 3-fold version of Chevet’s inequality (Lemma 3.3) and the Blaschke-Santal´oinequality (Lemma 3.1). The next theorem will be the consequence of the following two subsections.
Theorem B. Let n ∈ N and 1 ≤ p ≤ q ≤ r ≤ ∞. Then − min( 1 , 1 )−min( 1 , 1 )− 1 −1 1/n3 − min( 1 , 1 )−min( 1 , 1 )− 1 −1 p 2 q 2 r 3 n n n p 2 q 2 r n ≤ voln B`p ⊗π`q ⊗π`r .p,q,r n . 4.1. Lower bounds on the volume. Let us first we fix some more notation. For 1 ≤ p, q, r ≤ ∞ and n ∈ N, let n n n n Xπ := `p ⊗π `q ⊗π `r , where we suppress the parameters p, q, r that are clear from the context. For the correspond- n n n n ing norm, we write in short k·kπ instead of k·kXπ . Similarly, we define X := `p∗ ⊗ `q∗ ⊗ `r∗ n n and write k · k instead of k · kX . We denote the corresponding unit balls by Bπ and B n respectively. Any A ∈ X , we express in the form n X A = Ai,j,k ei ⊗ ej ⊗ ek. i,j,k=1 The main tool in proving a lower bound is the following estimate that compares the n3 injective norm k · k with the `1-norm in R .
8 n Proposition 4.1. Let n ∈ N, 1 ≤ p, q, r ≤ ∞ and assume A ∈ X . Then n 1 − min 1 , 1 −min 1 , 1 − 1 −1 X kAk ≥ n ( p 2 ) ( q 2 ) r |A | . 2 i,j,k i,j,k=1
n Proof. To prove this, we identify X with a space of operators. We have
n X kAk = max Ai,j,k xiyjzk . kxkp=kykq=kzkr=1 i,j,k=1 In what follows, ε, δ, η ∈ {−1, 1}n denote sign vectors. We divide the proof into three different cases, depending on which side of 2 the parameters p, q, r lie. − 1 Case 1: Assume that 2 ≤ p ≤ q. In this case we choose x = n p (ε1, . . . , εn), y := − 1 − 1 n q (δ1, . . . , δn), and z = n r (η1, . . . , ηn). Then we obtain
n n X − 1 − 1 − 1 X max Ai,j,kxiyjzk ≥ n p q r max Ai,j,k εiδjηk kxk =kyk =kzk =1 ε,δ,η∈{−1,1}n p q r i,j,k=1 i,j,k=1 n n − 1 − 1 − 1 X X = n p q r max Ai,j,k δjηk δ,η∈{−1,1}n i=1 j,k=1 n n − 1 − 1 − 1 X 1 X X ≥ n p q r max Ai,j,k δjηk . η∈{−1,1}n 2n i=1 δ∈{−1,1}n j,k=1 Applying Khintchine’s inequality and then H¨older’sinequality, we obtain
n n n n n 1/2 X 1 X X 1 X X X 2 max Ai,j,k δjηk ≥ √ max Ai,j,k ηk η∈{−1,1}n 2n 2 η∈{−1,1}n i=1 δ∈{−1,1}n j,k=1 i=1 j=1 k=1 n n n 1 X X X ≥ √ max Ai,j,k ηk . η∈{−1,1}n 2n i=1 j=1 k=1 Again, applying Khinchine’s inequality and then H¨older’s inequality,
n n n n n n X X X X X 1 X X max Ai,j,k ηk ≥ Ai,j,k ηk η∈{−1,1}n 2n i=1 j=1 k=1 i=1 j=1 η∈{−1,1}n k=1
n n n !1/2 1 X X X 2 ≥ √ |Ai,j,k| 2 i=1 j=1 k=1 n 1 X ≥ √ |Ai,j,k|. 2n i,j,k=1
n Case 2: Assume that p ≤ 2 ≤ q. In this case, we choose for x ∈ Bp the standard unit vectors e1, . . . , en and y, z as in the previous case. We get, again by using the inequalities of
9 Khintchine and H¨older, n n X − 1 − 1 X max Ai,j,k xiyjzk ≥ n q r max max Ai,j,k δjηk kxk =kyk =kzk =1 1≤i≤n δ,η∈{−1,1}n p q r i,j,k=1 j,k=1 n n − 1 − 1 X X = n q r max max Ai,j,k ηk 1≤i≤n η∈{−1,1}n j=1 k=1 n n − 1 − 1 −1 X X ≥ n q r max Ai,j,k ηk η∈{−1,1}n i,j=1 k=1 n n − 1 − 1 −1 X 1 X X ≥ n q r A η 2n i,j,k k i,j=1 η∈{−1,1}n k=1 n n !1/2 1 − 1 − 1 −1 X X 2 ≥ √ n q r |Ai,j,k| 2 i,j=1 k=1 n 1 − 1 − 1 − 3 X ≥ √ n q r 2 |Ai,j,k|. 2 i,j,k=1 n n Case 3: Assume that p ≤ q ≤ 2. Choose for x ∈ Bp and y ∈ Bq the standard unit vectors e1, . . . , en and z as in the previous two cases. We have n n X − 1 X max Ai,j,k xiyjzk ≥ n r max max Ai,j,k ηk kxk =kyk =kzk =1 1≤i,j≤n η∈{−1,1}n p q r j,k=1 k=1 n − 1 X = n r max |Ai,j,k| 1≤i,j≤n k=1 n − 1 −2 X ≥ n r |Ai,j,k|. i,j,k=1 This completes the proof of the proposition. As an immediate consequence of Lemma 4.1, we obtain a lower bound on the volume n n radius of the unit ball Bπ in Xπ . Corollary 4.2. Let n ∈ N and 1 ≤ p, q, r ≤ ∞. Then we have 1/n3 1 1 1 1 1 n − min( p , 2 )−min( q , 2 )− r −1 voln3 Bπ ≥ n . Proof. From Lemma 4.1, we obtain that
1 1 1 1 1 3 n min( p , 2 )+min( q , 2 )+ r +1 n B ⊆ 2 n B1 . Switching to the polar bodies implies
n 1 − min 1 , 1 −min 1 , 1 − 1 −1 n3 B ⊇ n ( p 2 ) ( q 2 ) r B . π 2 ∞ Taking volumes and the n3-rd root, the previous inclusion immediately gives 1/n3 1 1 1 1 1 n − min( p , 2 )−min( q , 2 )− r −1 voln3 Bπ ≥ n , which completes the proof.
10 4.2. Upper bounds on the volume. To compute the matching upper bound on the vol- n ume radius of Bπ , we will use Lemma 3.4. To be more precise, the idea is as follows. Using our extended version of Chevet’s inequality (see Lemma 3.3), we obtain a lower bound for n the volume of B from Lemma 3.4. We then use the Blaschke-Santal´oinequality (see Lemma n 3.1) to derive an upper bound on the volume of Bπ . The next proposition will be a consequence of the 3-fold Chevet inequality, where we n n n n apply Lemma 3.3 to the space X and choose (xi)i=1, (yj)j=1, (zk)k=1 to be the standard basis vectors. ∞ Proposition 4.3. Let n ∈ N and 1 ≤ p ≤ q ≤ r ≤ ∞. Let (εi,j,k)i,j,k=1 be a sequence of independent Bernoulli random variables. Then we have n 1 1 1 1 1 X max ∗ , +max ∗ , + ∗ −1 E εi,j,k ei ⊗ ej ⊗ ek p,q,r n p 2 q 2 r . (4.1) . i,j,k=1 Proof. First, recall that the Rademacher average is smaller than the Gaussian average (see Lemma 3.5) and so it is enough to prove inequality (4.1) with Gaussian random variables. In order to do that, recall the well known fact that, for all 1 ≤ α < ∞, and standard Gaussian random variables g1, . . . , gn, n X 1/α E giei α n . n i=1 `α n Also, it is known that if we consider the standard unit vectors e1, . . . , en in `α, then we have